Symphony: Learning Realistic and Diverse Agents for Autonomous Driving Simulation

ICRA 22 2022 35.3 method, application

TLDR

Symphony combines beam search and hierarchical goals to learn realistic and diverse driving agents for autonomous simulation.

Reasoning

The paper addresses mode collapse and improves realism via beam search and discriminator, using real-world datasets. However, it focuses narrowly on driving simulation and does not claim a general world model, limiting relevance to the specified keywords.

Read-first score

Read-first score 35.3, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 6.

Methodology quality 25%

Screens visible abstract and analysis fields for experiment, dataset, baseline, metric, and limitation evidence. markers=baseline,dataset,experiment

Recency 8%

56.5

Uses a gentle age decay so recent papers surface without erasing older foundations. 2022

Reproducibility 25%

Screens links and visible text for paper, code, dataset, artifact, and repository signals. pdf=True; code=False; dataset=False; markers=dataset

Topical relevance 42%

8.6

Uses existing LLM keyword relevance scores normalized to 0-100. world model,world simulator,generative world model,interactive world model,video world model,world dynamics prediction,model-based reinforcement learning world model

Field roles

Methodology anchor

Rank sensitivity

Stability: volatile; rank range: 66.

Keyword Scores

world simulator

world model

generative world model

interactive world model

world dynamics prediction

video world model

model-based reinforcement learning world model

Deep Analysis

Innovations

Combining conventional policies with a parallel beam search that prunes branches unfavorably evaluated by a discriminator to improve realism.
Hierarchical approach factoring agent behavior into goal generation and goal conditioning to preserve diversity during adversarial training and beam search.
Overall Symphony framework that achieves both realistic and diverse agent behavior for autonomous driving simulation.

Methodology

Symphony combines conventional policies with a parallel beam search that refines policies on the fly by pruning branches unfavorably evaluated by a discriminator. To prevent diversity loss from pruning, it uses a hierarchical approach that factors agent behavior into goal generation and goal conditioning, ensuring diversity is maintained. The method is trained using learning from demonstration on trajectory data from proprietary and open Waymo datasets.

Key Results

Experiments on both proprietary and open Waymo datasets show that Symphony agents learn more realistic and diverse behavior than several baselines.

Resources

Paper